British News Agency

Sovereign AI refers to a country’s ability to develop, host, deploy, and manage AI systems produced within its own borders, rather than depending on foreign systems or cloud jurisdictions.

In a report published in June, the European Parliament acknowledged that Europe is “currently heavily dependent on foreign technologies,” particularly American technologies that prevent the bloc from having its own tech champions. The report noted that this dependence is “likely to continue” due to the United States’ recent $500 billion (€432.9 billion) investment in domestic AI.

The EU stated that it must invest in research and the development of new systems to regain its competitive edge.

Germany became the latest country to announce its own AI initiative, called Sovereign Open-Source Foundation Models (SOOFI).

According to the German government, SOOFI is an initiative to build an open-source “advanced artificial intelligence” foundation model that can be adapted by other companies developing AI products.

The German government stated that the technology will be used for highly complex tasks, such as AI-controlled robots.

Professor Wolfgang Nejdl of Leibniz University Hannover, one of the universities participating in the project, said: “With SOOFI, we are laying the foundations for a new generation of sovereign, powerful, and entirely European AI models.”

Telecommunications companies Deutsche Telekom and T-Systems said they aim for SOOFI to have 100 billion parameters — the settings that regulate a model’s behavior.

Both companies are providing technical support for the large language model at one of the AI factories. Deutsche Telekom will use around 130 NVIDIA chips and more than 1,000 graphics processing units (GPUs), which will be ready for use by next March, to train the model.

In September, a Swiss AI startup launched Apertus, the country’s first multilingual language model.

Apertus — meaning “open” in Latin — allows researchers, professionals, and the public to customize the model according to their specific needs.

The developers state that everything related to the model — including training architecture, datasets, source code, and model weights (the parameters that teach an LLM how to interpret data) — is openly available.